A wider state, not wider arithmetic, for single-precision time integration

Michael F. Barad

Home/Archive/arXiv

arXiv · 2611.05365

A wider state, not wider arithmetic, for single-precision time integration

Michael F. Barad

View original source ↗Read source document ↗Export metadata ↓

Abstract

Single precision (FP32) halves the memory and data movement of double precision (FP64), and on accelerators is often several times faster. Its rounding errors are harmless except in the stored state. A modified equation for Runge-Kutta methods in Shu-Osher form and linear multistep methods separates the rounding of the stored state, a forcing proportional to $1/Δt$, from truncation and from right-hand-side rounding, which sets a $Δt$-independent floor. We bound this forcing for strong-stability-preserving schemes and identify two mechanisms that keep its sign fixed, so it accumulates as a bias: past a turnaround step, refining the step increases the error, in FP64 as in FP32. For a model oscillator the analysis predicts the floor and the FP32-to-FP64 ratios of turnaround step and minimum error within 30%. Our remedy needs no FP64 arithmetic: each state variable is stored as the unevaluated sum of two FP32 numbers, updated with error-free transformations, while every right-hand-side evaluation stays in FP32. Unlike compensated summation, it needs no rewriting of the scheme. In model problems from geophysical flow, mechanical dynamics and celestial mechanics, with five Runge-Kutta and multistep schemes, it restores design-order convergence to the floor. In large-eddy simulations of a high-lift airplane wing on three refined grids, up to 544 million cells, with time steps $Δt$, $Δt/2$ and $Δt/4$, FP32 overpredicts the lift by 2.1, 4.5 and 7.7%, while the two-float lift agrees with FP64 within 0.3% on every grid, and on graphics processing units the production solver runs 1.6 to 2.7 times faster than FP64.

Explore related subjects

math.NA cs.NA

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Michael F. Barad. 2026-11-03. A wider state, not wider arithmetic, for single-precision time integration. https://arxiv.org/abs/2611.05365

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Extracting Dual Analytic Geometries of Linear Transformations to Achieve Efficient Computation

We generalize the Calderon-Zygmund methodology for operator analysis and the Rokhlin-Greengard fast multipole algorithms, in which the domain and range of an operator are jointly organized. Our methods analyze unstructured kernels by uncovering the joint hidden geometries of their domain and range. The main tool is the Questionnaire algorithm, an iterative procedure that constructs adaptive bi-hierarchical partition trees that reveal latent multiscale organization and local low-rank structure in the kernel. When the geometry is unknown, the iteration starts from any initial affinity, even a random one, and, with a stochastic flip perturbation and a spectral space-filling curve that fixes the left-right order of every tree, reorganizes the matrix into an image of the operator's "phase space": the joint space of its sources and sensors, in duality, as positions and frequencies are for the Fourier transform. This ordering gives the phase space an analytic structure on which any classical transform can be carried out; the resulting compression is direct evidence of that structure. Guided by these geometries, we obtain efficient factorizations and fast matrix-vector products in two ways: (1) the Butterfly algorithm, which exploits the learned bi-hierarchical localized low-rank structure; and (2) the adaptive eGHWT, which selects best tilings in space and sequency from all levels of generalized Haar-Walsh wavelet packets. We call the resulting algorithms the Questionnaire Factorization and Fast Transform (QFFT). Unlike classical approaches requiring prior knowledge of the geometry, QFFT is fully data-driven and applies to matrices from irregular or unknown distributions. We demonstrate it on heterogeneous oscillatory operators and on orthogonal eigenvector expansions with high multiplicity and random sampling, reducing storage from O(N^2) to O(N log N) and enabling fast, scalable computation.

math.NA↗

arXiv2026

Decoupled energy-stable Runge-Kutta schemes of arbitrary order for the anisotropic phase-field dendritic crystal growth model

This paper introduces a novel and general framework for constructing arbitrary-order, energy-stable, and decoupled schemes for the anisotropic phase-field dendritic crystal growth model. The proposed framework combines a time-dependent auxiliary variable, an algebraically stable Runge-Kutta method, and a matrix diagonalization technique. This unique combination yields a scheme that is proven to be unconditionally energy-dissipative while requiring only the solution of independent, constant-coefficient elliptic equations at each time step, which can be solved in parallel. The auxiliary variable is recovered from a uniquely solvable $q\times q$ algebraic system. We provide a rigorous proof of $q$th-order convergence in time for a $q$-stage Runge-Kutta method within a fixed Fourier-Galerkin space. Numerical experiments in 2D and 3D not only confirm the theoretical convergence and energy stability but also demonstrate the efficiency of the decoupled approach and provide a detailed investigation of key parameters' effects on dendritic morphology.

math.NA↗

arXiv2026

Automated Numerical Stability Analysis of Deep Learning Operators

Finite-precision arithmetic unavoidably introduces numerical approximation errors. Numerical computations may use insufficient precision or an improper formulation, which leads to numerical instability. In this paper, we introduce a unified software tool for stochastic numerical validation of deep-learning operators. The tool follows CESTAC on supported exposed operations and uses an operator-level data-perturbation approximation for GEMM-like kernels. Our developed software not only enables numerical validation with a single computation pass but also detects the sources of numerical instability and provides numerical stability monitoring during deep learning training and inference. We verified its effectiveness on the detection of polluted operators with injected numerical instabilities across various tasks. We believe that our developed method and tools provide valuable insights into developing numerically stable computing kernels, which are particularly critical for numerically stable and efficient deep learning training and inference.

math.NA↗